roadrage10

gohumanize__gohumanize-open-humanizer

README

License: apache-2.0

How often it rewrites

A humanizer can fail by handing its input back almost unchanged. On 300 held-out passages, three tries each at temperature 0.9:

Table

Modern passages returned nearly unchanged (more than 90% of the words kept), from 59 articles never seen in training	17.0%
Book passages returned nearly unchanged	13.3%
Modern rewrites that lose a number	28.5%
Invented datelines, footnote numbers or links, of 900 outputs	0

The training pairs were built so the model learns to restructure sentences, not just swap words, and press-release datelines and footnote numbers were removed from the targets so the model does not learn to invent them. The write-up explains how.

Use

Chat format, thinking disabled. The system prompt below is the one used in training.

python

from transformers import AutoTokenizer, AutoModelForCausalLMimport torch repo = "gohumanize/gohumanize-open-humanizer"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto") SYSTEM = ("Rewrite the following text so that it reads as if a person wrote it: varied sentence "          "length, concrete wording, natural rhythm, no filler transitions. Keep the meaning, the "          "facts and the order of ideas. Return only the rewritten text.")messages = [{"role": "system", "content": SYSTEM},            {"role": "user", "content": "It is worth noting that the committee ultimately reached a consensus."}]ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)out = model.generate(ids, max_new_tokens=600, temperature=0.9, top_p=0.9, do_sample=True)print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

On modern text, a single try still comes back nearly unchanged about one time in six. Generating a few tries and keeping the one that changed most (the MCP server and Python client do this for you) takes that close to zero.

Run it locally without a GPU using the GGUF builds in gguf/:

bash

ollama run hf.co/gohumanize/gohumanize-open-humanizer:Q4_K_M

Also available: an MCP server (npx gohumanize-open-humanizer-mcp), a Python client (pip install gohumanize-open-humanizer) and a browser demo at https://gohumanize.ai/open-model.

Training

Table

Base model	Qwen/Qwen3-4B (Apache-2.0)
Method	Full fine-tune with Unsloth: every weight updated in bf16, 8-bit AdamW, gradient checkpointing
Data	2,957 train / 300 test pairs, see the dataset card
Loss	on the assistant (human target) tokens only
Epochs, LR, batch	2 epochs, 2e-5 cosine, effective batch 16, max 1,024 tokens, seed 13
Hardware	1x NVIDIA H100 (80 GB) on Modal, 11 minutes of training
Tracking	Weights & Biases run

Evaluation

The untouched base model and the Open Humanizer on the 300 held-out passages (100 modern, 200 from books), against the human original.

Table with columns: Measure (300 held-out pairs), AI-styled input, Base Qwen3-4B, Open Humanizer, Human target
Measure (300 held-out pairs)	AI-styled input	Base Qwen3-4B	Open Humanizer
BERTScore F1 vs human (higher = closer meaning)	0.918	0.903	0.929
ROUGE-L vs human (higher = closer wording)	0.522	0.434	0.580
Names/capitalised tokens kept (recall)	0.601	0.610

These measures cannot see whether the model rewrote at all: an output identical to the AI-styled input scores well on meaning and wording. That is why the rewrite rate above is measured separately.

These measure how far the output moves from AI-styled prose towards the human target. They are not detector scores.

Limitations

Two thirds of the human targets are pre-1929 books and the rest is US government writing, so the model leans literary on one side and official on the other. Casual modern writing is the gap.
About three in ten rewrites of number-heavy text lose at least one figure. Some of that is rewording ("20 percent" as "a fifth"), some is loss. Check figures and other facts against the source.
Trained on 80 to 300 word passages; rewrite long documents paragraph by paragraph.
English only.
Educational demo; no claims about AI detectors.

Links

Table with columns: Resource, Link
Resource	Link
Project page and browser demo	gohumanize.ai/open-model
GoHumanize (the product this research comes from)	gohumanize.ai
Model weights and GGUF builds	gohumanize/gohumanize-open-humanizer
Dataset, 3,257 pairs (CC-BY 4.0)	gohumanize/gohumanize-open-humanizer-dataset
Code and full pipeline

Licence and citation

Apache-2.0. Cite as:

GoHumanize team (2026). GoHumanize Open Humanizer: an open text-humanization model, dataset and pipeline built from public-domain data. Version 0.2.0. Zenodo. https://doi.org/10.5281/zenodo.22843083

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

roadrage10

Model Tree

Base

Qwen/Qwen3-4B

Quantized

this model

Input Modalities

Text

Output Modalities